1 //---------------------------------------------------------------------------//
2 // Copyright (c) 2013 Kyle Lutz <kyle.r.lutz@gmail.com>
3 //
4 // Distributed under the Boost Software License, Version 1.0
5 // See accompanying file LICENSE_1_0.txt or copy at
6 // http://www.boost.org/LICENSE_1_0.txt
7 //
8 // See http://boostorg.github.com/compute for more information.
9 //---------------------------------------------------------------------------//
10
11 #ifndef BOOST_COMPUTE_ALGORITHM_MERGE_HPP
12 #define BOOST_COMPUTE_ALGORITHM_MERGE_HPP
13
14 #include <boost/static_assert.hpp>
15
16 #include <boost/compute/system.hpp>
17 #include <boost/compute/command_queue.hpp>
18 #include <boost/compute/algorithm/copy.hpp>
19 #include <boost/compute/algorithm/detail/merge_with_merge_path.hpp>
20 #include <boost/compute/algorithm/detail/serial_merge.hpp>
21 #include <boost/compute/detail/iterator_range_size.hpp>
22 #include <boost/compute/detail/parameter_cache.hpp>
23 #include <boost/compute/type_traits/is_device_iterator.hpp>
24
25 namespace boost {
26 namespace compute {
27
28 /// Merges the sorted values in the range [\p first1, \p last1) with the sorted
29 /// values in the range [\p first2, last2) and stores the result in the range
30 /// beginning at \p result. Values are compared using the \p comp function. If
31 /// no comparision function is given, \c less is used.
32 ///
33 /// \param first1 first element in the first range to merge
34 /// \param last1 last element in the first range to merge
35 /// \param first2 first element in the second range to merge
36 /// \param last2 last element in the second range to merge
37 /// \param result first element in the result range
38 /// \param comp comparison function (by default \c less)
39 /// \param queue command queue to perform the operation
40 ///
41 /// \return \c OutputIterator to the end of the result range
42 ///
43 /// Space complexity: \Omega(distance(\p first1, \p last1) + distance(\p first2, \p last2))
44 ///
45 /// \see inplace_merge()
46 template<class InputIterator1,
47 class InputIterator2,
48 class OutputIterator,
49 class Compare>
merge(InputIterator1 first1,InputIterator1 last1,InputIterator2 first2,InputIterator2 last2,OutputIterator result,Compare comp,command_queue & queue=system::default_queue ())50 inline OutputIterator merge(InputIterator1 first1,
51 InputIterator1 last1,
52 InputIterator2 first2,
53 InputIterator2 last2,
54 OutputIterator result,
55 Compare comp,
56 command_queue &queue = system::default_queue())
57 {
58 BOOST_STATIC_ASSERT(is_device_iterator<InputIterator1>::value);
59 BOOST_STATIC_ASSERT(is_device_iterator<InputIterator2>::value);
60 BOOST_STATIC_ASSERT(is_device_iterator<OutputIterator>::value);
61 typedef typename std::iterator_traits<InputIterator1>::value_type input1_type;
62 typedef typename std::iterator_traits<InputIterator2>::value_type input2_type;
63 typedef typename std::iterator_traits<OutputIterator>::value_type output_type;
64
65 const device &device = queue.get_device();
66
67 std::string cache_key =
68 std::string("__boost_merge_") + type_name<input1_type>() + "_"
69 + type_name<input2_type>() + "_" + type_name<output_type>();
70 boost::shared_ptr<detail::parameter_cache> parameters =
71 detail::parameter_cache::get_global_cache(device);
72
73 // default serial merge threshold depends on device type
74 size_t default_serial_merge_threshold = 32768;
75 if(device.type() & device::gpu) {
76 default_serial_merge_threshold = 2048;
77 }
78
79 // loading serial merge threshold parameter
80 const size_t serial_merge_threshold =
81 parameters->get(cache_key, "serial_merge_threshold",
82 static_cast<uint_>(default_serial_merge_threshold));
83
84 // choosing merge algorithm
85 const size_t total_count =
86 detail::iterator_range_size(first1, last1)
87 + detail::iterator_range_size(first2, last2);
88 // for small inputs serial merge turns out to outperform
89 // merge with merge path algorithm
90 if(total_count <= serial_merge_threshold){
91 return detail::serial_merge(first1, last1, first2, last2, result, comp, queue);
92 }
93 return detail::merge_with_merge_path(first1, last1, first2, last2, result, comp, queue);
94 }
95
96 /// \overload
97 template<class InputIterator1, class InputIterator2, class OutputIterator>
merge(InputIterator1 first1,InputIterator1 last1,InputIterator2 first2,InputIterator2 last2,OutputIterator result,command_queue & queue=system::default_queue ())98 inline OutputIterator merge(InputIterator1 first1,
99 InputIterator1 last1,
100 InputIterator2 first2,
101 InputIterator2 last2,
102 OutputIterator result,
103 command_queue &queue = system::default_queue())
104 {
105 BOOST_STATIC_ASSERT(is_device_iterator<InputIterator1>::value);
106 BOOST_STATIC_ASSERT(is_device_iterator<InputIterator2>::value);
107 BOOST_STATIC_ASSERT(is_device_iterator<OutputIterator>::value);
108 typedef typename std::iterator_traits<InputIterator1>::value_type value_type;
109 less<value_type> less_than;
110 return merge(first1, last1, first2, last2, result, less_than, queue);
111 }
112
113 } // end compute namespace
114 } // end boost namespace
115
116 #endif // BOOST_COMPUTE_ALGORITHM_MERGE_HPP
117